Increased number of regional lymph nodes containing metastases

Transcription

1 ORIGINAL ARTICLES for Esophageal Cancer Nabil P. Rizk, MD,* Hemant Ishwaran, PhD, Thomas W. Rice, MD, Long-Qi Chen, MD, Paul H. Schipper, MD, Kenneth A. Kesler, MD, Simon Law, MD, Toni E. M. R. Lerut, MD, PhD,** Carolyn E. Reed, MD, Jarmo A. Salo, MD, Walter J. Scott, MD, Wayne L. Hofstetter, MD, Thomas J. Watson, MD, Mark S. Allen, MD,*** Valerie W. Rusch, MD,* and Eugene H. Blackstone, MD Objective: Using Worldwide Esophageal Cancer Collaboration data, we sought to (1) characterize the relationship between survival and extent of lymphadenectomy, and (2) from this, define optimum lymphadenectomy. Summary Background Data: What constitutes optimum lymphadenectomy to maximize survival is controversial because of variable goals, analytic methodology, and generalizability of the underpinning data. Methods: A total of 4627 patients who had esophagectomy alone for esophageal cancer were identified from the Worldwide Esophageal Cancer Collaboration database. Patient-specific risk-adjusted survival was estimated using random survival forests. Risk-adjusted 5-year survival was averaged for each number of lymph nodes resected and its relation to cancer characteristics explored. Optimum number of nodes that should be resected to maximize 5-year survival was determined by random forest multivariable regression. Results: For pn0m0 moderately and poorly differentiated cancers, and all node-positive (pn ) cancers, 5-year survival improved with increasing extent of lymphadenectomy. In pn0m0 cancers, no optimum lymphadenectomy was defined for ptis; optimum lymphadenectomy was 10 to 12 nodes for p, 15 to 22 for p, and 31 to 42 for p, depending on histopathologic cell type. In pn M0 cancers and 1 to 6 nodes positive, optimum lymphadenectomy was 10 for p, 15 for p, and 29 to 50 for p. Conclusions: Greater extent of lymphadenectomy was associated with increased survival for all patients with esophageal cancer except at the extremes (TisN0M0 and 7 regional lymph nodes positive for cancer) and well-differentiated pn0m0 cancer. Maximum 5-year survival is modulated by T classification: resecting 10 nodes for p, 20 for p, and 30 for p is recommended. (Ann Surg 2010;251: 46 50) From the *Memorial Sloan-Kettering Cancer Center, New York, NY; Cleveland Clinic, Cleveland, OH; West China Hospital of Sichuan University, Chengdu, Sichuan, People s Republic of China; Oregon Health and Science University, Portland, OR; Indiana University, Indianapolis, IN; Queen Mary Hospital, University of Hong Kong Medical Center, Hong Kong, China; **University Hospital Leuven, Leuven, Belgium; Medical University of South Carolina, Charleston, SC; Helsinki University Central Hospital, Helsinki, Finland; Fox Chase Cancer Center, Philadelphia, PA; University of Texas MD Anderson Cancer Center, Houston, TX; University of Rochester, Rochester, NY; and the ***Mayo Clinic, Rochester, MN. Supported by The Cleveland Clinic Foundation, the Daniel and Karen Lee Endowed Chair in Thoracic Surgery (to T.W.R.), and the Kenneth Gee and Paula Shaw, PhD, Chair in Heart Research (to E.H.B.). Reprints: Thomas W. Rice, MD, Cleveland Clinic, Department of Thoracic and Cardiovascular Surgery, 9500 Euclid Ave./Mail Stop J4-1, Cleveland, Ohio ricet@ccf.org. Supplemental digital content is available for this article. Direct URL citations appear in the printed text, and links to the digital files are available in the HTML text of this article on the journal s Web site ( Copyright 2009 by Lippincott Williams & Wilkins ISSN: /10/ DOI: /SLA.0b013e3181b2f6ee 46 Increased number of regional lymph nodes containing metastases predicts decreased survival following esophagectomy for cancer, 1 6 and increased extent of lymphadenectomy is associated with improved survival. 1,7 12 Therefore, lymphadenectomy of some extent is required. However, what constitutes optimum lymphadenectomy to maximize survival is controversial. 1,8 11 Using Worldwide Esophageal Cancer Collaboration (WECC) data, 2 we sought to (1) characterize the relationship between survival and extent of lymphadenectomy, and from this, (2) define optimum lymphadenectomy. METHODS Worldwide Esophageal Cancer Collaboration WECC is a consortium of institutions (Appendix 1, Supplemental Digital Content 1, available at: that has contributed deidentified patient data on esophagectomy for cancer. All datasets were approved for research by each site s institutional review board, and data use agreements were executed when required. The entire project was approved by the Case Cancer Institutional Review Board of Case Western Reserve University. Patients A total of 4627 patients underwent esophagectomy alone (no pre- or postoperative adjuvant therapy) for esophageal cancer and had follow-up for all-cause mortality. 2 Characteristics of these patients and their cancers are presented in Table 1, including number of regional lymph nodes resected (hereafter termed nodes resected ) and number of regional lymph nodes involved with cancer (hereafter termed nodes positive, pn ). Although the analysis examined number of nodes resected and nodes positive in a continuous fashion, for illustration, number of nodes positive is separated into 4 groups: (1) no nodes positive, (2) 1 to 2 nodes positive, (3) 3 to 6 nodes positive, and (4) 7 or more nodes positive. Data Analysis Multivariable Risk-Adjusted Survival Random survival forests (RSF) methodology 13,14 using the randomsurvivalforest R-package under its default settings 15 was employed to calculate a survival curve for each patient, adjusted for 45 variables (Appendix 2, Supplemental Digital Content 2, available at: These included TNM classifications, number of nodes resected, number of nodes positive (N), histopathologic cell type, histologic grade (G), location and length of cancer, residual cancer (R), patient demographics (age, race, gender), country, institution, and year of surgery. For this, RSF analysis generated a forest of 1000 random bootstrap survival trees, each grown using log-rank splitting. Missing nonanatomic and demographic data were imputed using a dynamic tree method. 14 On average, each bootstrap survival tree was constructed from 63% of the patients (bootstrapped data); the remaining 37% was referred to as out-of-bag (OOB) data. These OOB data and the corresponding bootstrap survival tree were used to generate an OOB survival curve for each patient in the OOB dataset. Growing 1000 trees yielded approximately 370 OOB sur- Annals of Surgery Volume 251, Number 1, January 2010

2 Annals of Surgery Volume 251, Number 1, January 2010 TABLE 1. Patient and Esophageal Cancer Characteristics Characteristic n* Mean SD or No. (% of n) Demography Age (yr) Men (77) Race 3587 White 2339 (65) Asian 1168 (33) Other 80 (2.2) Cancer location 4344 Upper third 177 (4.1) Middle third 1172 (27) Lower third 2995 (69) Cancer length pt 4609 is 335 (7.3) (23) (16) (51) (3.3) pn (56) (44) Number of nodes positive (57) 900 (20) 599 (13) 7 or more 424 (9.4) Number of nodes resected (1.1) (25) (19) (14) (11) (8.6) (5.6) (3.9) (2.9) (8.4) pm (92) (7.8) Histopathologic cell type 4595 Adenocarcinoma 2775 (60) Squamous 1834 (40) Undifferentiated 7 (0.15) Histologic grade 3816 G (32) G (33) G (35) G4 7 (0.18) Resection margins 4123 R (87) R1 434 (11) R2 117 (3.0) *Number of cases with values available. pa in 262 and pb in 244 among 506 in whom this distinction was made. pm1a in 104 and pm1b in 122 of 226 in whom this distinction was made. In 21, both cell types existed. vival curves for each patient, which were averaged to yield a risk-adjusted OOB ensemble survival curve. Risk-adjusted OOB ensemble 5-year survival was obtained by extracting the survival value at 5 years from these curves. 16 Averaged Risk-Adjusted 5-Year Survival Risk-adjusted OOB ensemble 5-year survival was averaged for all patients with a given number of nodes resected to yield averaged risk-adjusted 5-year survival. Because histologic grade and histopathologic cell type were previously found to predict survival for pn0m0 (node negative, no distant metastases) cancers, but not for pn M0 (node positive, no distant metastases) cancers, 16 separate analyses were performed for pn0m0 and pn M0 cancers. For pn0m0 data, averaging was stratified by histologic grade and histopathologic cell type within T. For pn M0 data, averaging was stratified by number of nodes positive within T. Random forests multivariable regression (RF-R) was used to determine optimum number of nodes that should be resected (lymphadenectomy) to maximize 5-year survival. 17 For pn0m0 cancers, number of nodes resected, histologic grade, and location were independent variables, with risk-adjusted 5-year survival the response. For each combination of T and histopathologic cell type, regression was performed for those with number of nodes resected greater than or equal to a fixed cutoff value ranging from 0 to 50. Only combinations with more than 15 patients were analyzed. Each analysis generated 1000 random bootstrap regression trees. Computations were performed with the randomforest R-package under its default settings. 18 For every cutoff value, RF regression yielded a variable importance (VIMP) value for each independent variable. VIMP measures predictiveness of a variable, adjusting for all other variables; it is the change in mean-square error of predicted 5-year survival after the variable is removed from the analysis. VIMP was standardized by dividing it by the variance of risk-adjusted 5-year survival and multiplying by 100, yielding values from 100% to 100%. A standardized VIMP 0 signifies a predictive variable; however, a threshold of 5% was used for this study to define predictiveness. This threshold accounts for Monte Carlo and sampling variability. For each stratification, optimum number of nodes resected was determined by the value at which standardized VIMP first dropped below 5%. For pn M0 cancers, a similar RF-based strategy was used. Data were stratified by number of nodes positive within T: 1 or 2, 3 to 6, and 7 or more. Number of nodes resected and number of nodes positive were used as independent variables (the latter was included to allow for more precise prediction, even though data were stratified into node-positive groupings). Standardized VIMP was calculated for number of resected nodes, stratified by number of nodes positive within T classification. RESULTS Survival and Extent of Lymphadenectomy pn0m0 Cancers For patients with ptis cancers, regardless of histopathologic cell type, survival was excellent and not associated with extent of lymphadenectomy (Fig. 1A). For patients with N0M0 cancers, survival was unrelated to extent of lymphadenectomy for G1 cancers (Fig. 1A), but increased with more extensive lymphadenectomy for G2/G3 cancers (Fig. 1B). For patients with pn0m0 and pn0m0 cancers, there were few G1 cancers with extensive lymphadenectomy (Fig. 1A), but the limited data suggest survival was unrelated to extent of lymphade Lippincott Williams & Wilkins 47

3 Annals of Surgery Volume 251, Number 1, January 2010 Rizk et al FIGURE 2. Coplot of averaged risk-adjusted 5-year survival after esophagectomy for pn M0 cancers according to number of nodes resected, T classification, and number of nodes positive. Solid lines are loess smoothed curves. pnⴙm0 Cancers For patients with pn M0 cancers and 1 to 6 nodes positive, survival increased with extent of lymphadenectomy for all T classifications (Fig. 2). For patients with 7 or more nodes positive, survival increased, albeit minimally, with extent of lymphadenectomy for and cancers. Paucity of p cancers with 7 or more nodes positive precluded assessing the survival value of lymphadenectomy. pn0/pnⴙ Distinction FIGURE 1. Coplot of averaged risk-adjusted 5-year survival after esophagectomy for pn0m0 cancers according to number of nodes resected, pt classification, and histopathologic cell type. A, Raw data. Each dot represents an average for a node resected. Colors represent histologic grade (see key). B, G2/G3 cancers with survival displayed as smooth loess curves according to number of nodes resected. nectomy; however, for G2/G3 cancers, 5-year survival increased with extent of lymphadenectomy (Fig. 1B). For p N0M0 cancers, patients with minimal lymphadenectomy (Fig. 1B) have survival equivalent to patients with p N M0 cancers with moderate to extensive lymphadenectomy and 1 to 6 nodes positive (Fig. 2). For pn M0 cancers, for the same T classification, extensive lymphadenectomy yielded 5-year survival similar to minimal lymphadenectomy for one lower node-positive grouping (Table 2). For ptis, no optimum extent of lymphadenectomy could be identified (Fig. 3A). For p cancers, optimum lymphadenectomy was 10 for N0M0 adenocarcinomas and 12 for N0M0 squamous cell carcinomas (Fig. 3A), and 10 for 1 to 6 nodes positive (insufficient data were available for 7 or more nodes positive) (Fig. 3B). For p cancers, optimum lymphadenectomy was 15 for N0M0 adenocarcinomas and 22 for N0M0 squamous cell carcinomas (Fig. 3A), and TABLE 2. ptn Classifications With Similar Survival, Illustrating Survival Difference Between Minimal and Extensive Lymphadenectomy Extensive Lymphadenectomy pt Minimal Lymphadenectomy Nodes Positive Nodes Resected 5-yr Risk-Adjusted Survival (%) 5-yr Risk-Adjusted Survival (%) Nodes Resected Nodes Positive pt ⱕ5 ⱕ Lippincott Williams & Wilkins

4 Annals of Surgery Volume 251, Number 1, January 2010 extensive lymphadenectomy; we believe this represents stage purification. However, for pn cancers, improved survival with more extensive lymphadenectomy may relate to more accurate determination of number of positive nodes (stage purification), or therapeutic effect of removing micrometastases. A unique aspect of this study is that optimal lymphadenectomy was obtained using a new machine-learning technique. The technique is related to traditional recursive partitioning analysis, which has been used extensively in analysis of cancer survival data. 1,21 However, random forest (RF) analyses introduce randomization by bootstrapping the data to result in multiple (hundreds to thousands) of trees. What is generally found is that each bootstrapped tree is a different result. Thus, any given tree (or recursive partitioning analysis) is unstable. 22 But when a forest of such trees is summarized (aggregated), a stable result emerges. 14 Like all varieties of RF analysis, RSF analysis makes no a priori assumptions about survival as do model-based analyses, which may assume proportional hazards, a particular formulation of a risk factor model, and a specific form for interaction terms. Most importantly, because trees in the forest use recursive partitioning, RSF is able to identify complex interactions between variables and nonlinear relationships, 14,17 something that is difficult to do using parametric and semiparametric methods. VIMP was used to define optimal lymphadenectomy. Rather than examining goodness of fit or P values to test for a statistically significant effect, the present analysis focused on predictiveness for future patients. Because VIMP uses RF regression, it is nonparametric and therefore has the same advantages as RSF. Stratification of the VIMP analysis by T, histopathologic cell type, and N, and use of histologic grade, tumor location, and number of nodes resected and positive as independent variables, refined the determination of optimal lymphadenectomy. Use of nonanatomic tumor characteristics and demographic variables in addition to TNM was based on previously identifying these as crucial correlates of survival. 16 FIGURE 3. Standardized variable importance (VIMP) for a series of cutoff values for number of nodes resected according to T classification. A, pn0m0 cancers, stratified by histopathologic cell type. B, pn M0 cancers, stratified by number of nodes positive. 15 for 1 to 6 nodes positive (insufficient data were available for 7 or more nodes positive) (Fig. 3B). For p cancers, optimum lymphadenectomy was 31 for N0M0 adenocarcinomas and 42 for N0M0 squamous cell carcinomas (Fig. 3A), and 29 for 1 to 2 nodes positive, 50 for 3 to 6 nodes positive, and 28 for 7 or more nodes positive (Fig. 3B). DISCUSSION Principal Findings and Uniqueness of Analysis Based on worldwide data, extent of lymphadenectomy was either unassociated with or minimally increased survival for patients with extremes of esophageal cancer (TisN0M0 and 7 or more nodes positive) and those with well-differentiated pn0 cancer. Five-year survival improved with increasing extent of lymphadenectomy for all other pn0 and pn cancers. RF analysis identified optimum lymphadenectomy for each pt and confirmed the interplay of depth of tumor invasion (T) and N. 19,20 Whether the effect of lymphadenectomy is stage purification or therapy cannot be resolved completely from this study. For example, survival after minimal lymphadenectomy for pn0 cancers is equivalent to that of pn cancers with 1 to 6 positive nodes after moderate to Limitations Extent of lymphadenectomy did not occur according to a uniform protocol and so was highly variable. Similarly, there was no uniform protocol for pathologic review of the resection specimen. We exploited this heterogeneity to investigate optimal lymphadenectomy. Despite worldwide data, there was a paucity of cases at the extremes, such as well-differentiated pn0 and p/ with 7 or more nodes positive; this relates as much to biology as to the study. The measure of optimal lymphadenectomy was all-cause mortality, which includes a few noncancer deaths, but nevertheless, it is a reliable end point and the basis for most cancer staging. 23,24 We did not have morbidity information according to extent of lymphadenectomy, although hospital mortality was low. 2 The main problem is that each institution has a different method of counting the number of lymph nodes resected. Some pathology laboratories may not be as fastidious as others and therefore provide an artificially low count. Quest to Define For many investigators, the quest to define optimum lymphadenectomy has focused on discovering a single number for nodes resected. Evidence from our study indicates that an increasing extent of lymphadenectomy is associated with progressively increased survival that plateaus a nonlinear effect. There is no single number. Few have investigated the possibility that adequacy of lymphadenectomy-only depends on cancer characteristics what might be called an interaction effect. 1,8 Where it has, as in the present study, an important difference in optimum lymphadenectomy has been found that varies with cancer characteristics Lippincott Williams & Wilkins 49

5 Rizk et al Annals of Surgery Volume 251, Number 1, January 2010 Mechanics of analysis also has influenced defining optimum lymphadenectomy. Some have formally considered a nonlinear relationship, such as we have found, 1,8 others have assumed a linear effect, 10,11 and still others have arbitrarily grouped data as a form of acknowledging nonlinear effects. 9,12 In addition, the statistical criterion for defining an optimum lymphadenectomy has differed. Most have focused on prognosis, either maximizing survival or minimizing hazard; however, others have focused on goodness of fit to minimize classification error. 10,11 Generalizability of results plays a role. Many studies are from a single institution with a relatively small number of patients. 1,8,12 Others use epidemiologic databases not designed specifically for studying lymphadenectomy. 9,10 We and others 11 have used worldwide data from multiple institutions because this is an uncommon cancer, and its characteristics differ geographically. Finally, some have avoided number of nodes resected and simply compared fields of dissection (transhiatal, 2- and 3-field lymphadenectomy). 7 Extent of lymphadenectomy in such analyses is surgeon dependent and so is coarse and highly variable. All of these analyses have led to different recommendations for extent of lymphadenectomy. Rizk et al 1 used recursive partitioning analysis to study 336 esophagectomies from a single institution. They identified 18 nodes resected as the minimum necessary for accurate staging and for eliminating an effect of lymphadenectomy on survival. Altorki et al 8 used proportional hazards multivariable analysis of 264 esophagectomies from a single institution. Patients were stratified into 4 groups by number of nodes resected. Effect of lymphadenectomy on survival was lost after 25 nodes resected for early-stage disease and after 16 nodes in stage III and IV cancers. Greenstein et al 9 used both univariable log-rank tests of stratified Kaplan-Meier survival curves and multivariable proportional hazards regression with SEER data (1998 to 2003) from 972 esophagectomies. Patients were stratified by number of nodes resected. Survival progressively increased with increased extent of lymphadenectomy in pn0 cancer. They recommended resecting at least 18 lymph nodes. Schwarz and Smith 10 used multivariable proportional hazards analysis of SEER data (1970 to 2003) of 5620 esophagectomies. Thirty or more nodes resected was associated with best survival. Peyre et al 11 used logistic and Cox regression analysis of an international database of 2303 esophagectomies. Goodness of fit to survival was maximized with 23 to 29 nodes resected. Recommendations If there is uncertainty as to T and histopathologic grade, it is recommended that 30 or more nodes be resected to maximize 5-year survival. Preoperative assessment requires biopsy of the cancer to determine histopathologic cell type and histologic grade, and endoscopic esophageal ultrasound to determine ct, which is a reasonably accurate reflection of pt. 25,26 It is recommended that to maximize 5-year survival, a minimum of 10 nodes be resected for cancer, 20 nodes for cancer, and 30 or more nodes for cancers. True extent of lymphadenectomy is best assured when the surgeon separates the nodal material from the specimen, sending separate packets of nodal material to the pathology laboratory; 27 however, care must be exercised in recovering nodes surrounding the primary cancer so as not to obscure the radial resection margin. ACKNOWLEDGMENTS The authors thank Carolyn Apperson-Hansen, Elizabeth Lieber, Linda DiPaola, Peter Kisuule, Areatha Lockridge, Penny Houghtaling, Tess Parry, and Maureen Kidd for their helpful contributions to this study. REFERENCES 1. Rizk N, Venkatraman E, Park B, et al. The prognostic importance of the number of involved lymph nodes in esophageal cancer: implications for revisions of the American Joint Committee on Cancer staging system. J Thorac Cardiovasc Surg. 2006;132: Rice TW, Rusch VW, Apperson-Hansen C, et al. Worldwide esophageal cancer collaboration. Dis Esophagus. 2009;22: Wilson M, Rosato EL, Chojnacki KA, et al. Prognostic significance of lymph node metastases and ratio in esophageal cancer. J Surg Res. 2008;146: Mariette C, Piessen G, Briez N, et al. The number of metastatic lymph nodes and the ratio between metastatic and examined lymph nodes are independent prognostic factors in esophageal cancer regardless of neoadjuvant chemoradiation or lymphadenectomy extent. Ann Surg. 2008;247: Martin DJ, Church NG, Kennedy CW, et al. Does systematic 2-field lymphadenectomy for esophageal malignancy offer a survival advantage? Results from 178 consecutive patients. Dis Esophagus. 2008;21: Hofstetter W, Correa AM, Bekele N, et al. Proposed modification of nodal status in AJCC esophageal cancer staging system. Ann Thorac Surg. 2007; 84: ; discussion Kang CH, Kim YT, Jeon SH, et al. Lymphadenectomy extent is closely related to long-term survival in esophageal cancer. Eur J Cardiothorac Surg. 2007;31: Altorki NK, Zhou XK, Stiles B, et al. Total number of resected lymph nodes predicts survival in esophageal cancer. Ann Surg. 2008;248: Greenstein AJ, Litle VR, Swanson SJ, et al. Effect of the number of lymph nodes sampled on postoperative survival of lymph node-negative esophageal cancer. Cancer. 2008;112: Schwarz RE, Smith DD. Clinical impact of lymphadenectomy extent in resectable esophageal cancer. J Gastrointest Surg. 2007;11: ; discussion Peyre CG, Hagen JA, DeMeester SR, et al. The number of lymph nodes removed predicts survival in esophageal cancer: an international study on the impact of extent of surgical resection. Ann Surg. 2008;248: Bogoevski D, Onken F, Koenig A, et al. Is it time for a new TNM classification in esophageal carcinoma? Ann Surg. 2008;247: Ishwaran H, Kogalar UB. Random survival forests for R. R News. 2007;7: Ishwaran H, Kogalur UB, Blackstone EH, et al. Random survival forests. Ann Appl Stat. 2008;2: Ishwaran H, Kogalur UB. Random survival forest R package Available at: Ishwaran H, Blackstone EH, Apperson-Hansen C, et al. A novel data-driven approach to stage grouping of cancer: application to esophageal cancer. Stat Med In press. 17. Breiman L. Random forests. Mach Learn. 2001;45: Liaw A, Wiener M. Classification and regression by randomforest. R News. 2002;2: Rice TW, Zuccaro G Jr, Adelstein DJ, et al. Esophageal carcinoma: depth of tumor invasion is predictive of regional lymph node status. Ann Thorac Surg. 1998;65: Akiyama H, Tsurumaru M, Kawamura T, et al. Principles of surgical treatment for carcinoma of the esophagus: analysis of lymph node involvement. Ann Surg. 1981;194: Rice TW, Blackstone EH, Rybicki LA, et al. Refining esophageal cancer staging. J Thorac Cardiovasc Surg. 2003;125: Breiman L. Heuristics of instability and stabilization in model selection. Ann Stat. 1996;24: AJCC Cancer Staging Manual. 6th ed. New York, NY: Springer-Verlag; Ries LG, Young JL, Keel GE, et al, eds. SEER Survival Monograph: Cancer Survival Among Adults: US SEER Program, , Patient and Tumor Characteristics. Bethesda, MD: National Cancer Institute, SEER Program; NIH publication Rosch T. Endosonographic staging of esophageal cancer: a review of literature results. Gastrointest Endosc Clin N Am. 1995;5: Saunders HS, Wolfman NT, Ott DJ. Esophageal cancer. Radiologic staging. Radiol Clin N Am. 1997;35: Veeramachaneni NK, Zoole JB, Decker PA, et al. Lymph node analysis in esophageal resection: American College of Surgeons Oncology Group Z0060 trial. Ann Thorac Surg. 2008;86: ; discussion Lippincott Williams & Wilkins